Experience Training is a great way to improve the quality of search results by training the Search algorithm with feedback on its predictions. It is a critical tool for debugging backend issues and optimizing results for all search experiences.

Rejecting or modifying a prediction overrides that query immediately. All feedback (approving, rejecting, or changing) helps Search improve the relevancy of results over time.

Advanced Search Tier Features: Featured snippets and inferred filters, except on builtin.location, are only available on the Advanced Search tier. Check out the Search Tiers reference doc for details.

Experience Training allows you to train the Search algorithm in three areas: Featured Snippets, Inferred Filters, and Spell Checking.

To train your experience in the UI, navigate to the desired experience under Search and find the Training section. You can also find it via configuration as code (CaC).

Note: The thumbs up/thumbs down buttons on results cards in the frontend are not experience training. They do not directly change search ranking, retrain the model, or alter featured snippet selection. They only provide analytics on how end users view the quality of answers.

Navigating the Experience Training Tables

Each Experience Training screen shows a table of predictions made by an AI/ML model within Search, organized by search term for the last seven days. There's a row for every prediction for each search term, so a search term may have multiple predictions.

Features available in all three tables:

• Search bar: Search for specific search terms to see their training predictions.
• New vs. Completed: Toggle "Show completed" to see training predictions that have already been approved, rejected, or modified. Especially useful when debugging.
• Search Count: Number of searches for that term in the last 30 days. Sort by this column to see the most frequently searched terms.
• Last Searched: The most recent search time. The table is sorted by this column by default.
• Training Actions: In the right column, action buttons let you approve (checkmark), reject (x), or change (wand icon) training predictions.

Featured Snippets

On the Featured Snippets training screen, you'll see search terms for which the algorithm returned one or more potential featured snippets. Featured snippets extract a direct answer to a user query from unstructured data. See the Direct Answers unit for more detail.

The table also shows entity information — the entity (hyperlinked) and field the featured snippet came from, which is crucial context to determine whether a snippet is correct.

With featured snippets training, there are three possible actions:

• Reject: Suppress the snippet from results — it will never return for that search term.
• Approve: Train the algorithm and allow the snippet to display on verticals with the featured snippets prediction mode set to APPROVE_ONLY.
• Change: The modified snippet will show in future search results. Editing is not currently supported for snippets pulled from files.

Note: If the featured snippets prediction mode is set to APPROVE_ONLY, changing the underlying content of an approved snippet means that snippet is no longer approved. You'll need to re-approve it if you want it to appear with updated content.

Change Featured Snippets Modal

When you click the change wand button, a modal appears that allows you to approve, reject, or change featured snippets. If the algorithm identifies the incorrect portion of text from your entity, you can select a different portion of text to use as the featured snippet going forward. You can also change the entity the snippet pulls from via the Search Merchandiser.

When changing plain text featured snippets, click and drag your mouse across the text you want returned. The modal will automatically expand the highlighted section to include only full words.

When changing rich text featured snippets, enable the Rich Text toggle and select text by individually clicking paragraphs or sections in sequential order.

Inferred Filters

On the Inferred Filters training screen, you'll see search terms that had inferred filters applied. The Filters column shows the vertical the filter was applied to, the field, and the value to filter on. See the Searchable and Display Fields unit for more detail on inferred filters.

With inferred filters training, there are two possible actions: approve or change. Approving helps train the algorithm to better understand your brand. Changing overrides other algorithms to apply the new inferred filters.

By choosing to change filters, you can add, remove, or change the filters applied for a specific query. Each filter has a vertical, field, and value — those are what you can change.

Example: When a user searches "jobs in New York" the algorithm may apply Entity Type = Jobs and Location = York. You want the location filter to be "New York," not "York." Override this by selecting the Jobs vertical in the modal and editing the Location filter to be near New York.

Spell Checking

On the Spell Checking training screen, you can review suggested spell checks and approve or reject corrections. These are based solely on search terms entered. If you reject a correction, the algorithm will no longer surface that correction when that search term is run for that experience.

Troubleshooting Experience Training Overrides

The system allows you to approve or reject multiple overrides for the same input, which can cause unexpected behavior. Before approving or rejecting a suggestion, check whether there are already overlapping completed overrides for that query. The system attempts to pick the latest override to apply, but having multiple overrides for the same search term can confuse the algorithm.

General Troubleshooting Steps

1. Navigate to Search > [Your Experience] > Featured Snippets / Inferred Filters / Spell Checking — whichever type you are troubleshooting.
2. Use the search bar to filter to the search term you are troubleshooting.
3. Click Show Completed to see overrides that have already been approved, rejected, or modified.
4. Check for any entries for the relevant vertical:
   • If the desired result was rejected, it will not be applied for this search query. To apply it, undo the completed override.
   • If a different result was approved, the desired one will not be applied. Undo the incorrect override, toggle "Show completed" off, then approve the desired result by clicking the green checkmark.
   • If there are multiple active overrides, undo the undesired ones so only one active override remains.

No Inferred Filter / Featured Snippet / Spell Check Is Applied

If you reject a suggestion, the algorithm does not automatically select the next-best option. Follow the steps above to check whether a search term has a rejected suggestion but no approved one, then approve the desired result. You may need to wait for the system to suggest a new override after the relevant search query is run again.

My Featured Snippet Override Is No Longer Working

Check whether the content of the entity the snippet is extracted from has changed. If the content changed and Search can no longer find the snippet in the body, the override becomes stale and will not match.

To re-create the override:

1. Undo the existing override.
2. Approve a more recent featured snippet with the correct body.